Software Engineering Courses (SECourses)•15mo ago

<@205854764540362752> this is what kohya said about <@1236274587520598136> method

this is what kohya said about method

Maxivi · 2024-10-06T09:51:18.070Z

this is what kohya said about method

@Furkan Gözükara SECourses this is what kohya said about @dxqb method

MMaxivi <@205854764540362752> this is what kohya said about <@1236274587520598136> metho...

dxqb•10/6/24, 10:29 AM

He is right, it is very similar. You could have exactly the same effect by pregenerating and therefore make this possible for full finetune with no additional vram, too. You'd have to synchronize timesteps and seed(!) between the pregenerated data and the training though, which is the difference to the current "Prior Preservation" feature. I did implement this for full finetune also, but without pregeneration so it needs a lot of VRAM for two full models, the student and the trainer model. Feel free to forward my contact.

dxqb•10/6/24, 10:39 AM

On the second comment though, it's not only activation. In this sample, dreambooth-training, Clinton would look like the left person without regularization.

dxqb•10/6/24, 10:39 AM

the left person in the right sample, I mean

MMaxivi <@205854764540362752> this is what kohya said about <@1236274587520598136> metho...